PROJECT SUMMARY/ABSTRACT
The proposed study will develop a screening tool using electronic health record data that predicts the risk of
ED return and associated morbidity or mortality to support safe and appropriate dispositions in the ED for
patients with the novel coronavirus disease-2019 (COVID-19). Due to the challenges of COVID-19, with highly
variable symptoms, the paucity of existing research, and strains on ED capacity, emergency clinicians must
make rapid clinical decisions with limited information. Moreover, in the ED, patients often present for evaluation
early on during the course of their illness, which is when the clinical trajectory for COVID-19 is most volatile
and the risk for subsequent decompensation is highest. Using predictive modeling with natural language
processing (NLP) and machine learning (ML) techniques can leverage the data-rich environment of the ED to
improve the quality of care delivered to patients with COVID-19.
This study directly addresses priorities highlighted in PA-17-246 by bringing research evidence to clinical
practice through the development and evaluation a health IT solution that combines the use of NLP with a
decision support tool to turn unstructured clinical data into knowledge that can be applied to practice.
Developing and operationalizing the proposed COVID-19 ED return screening tool (CERST) can help ED
clinicians avoid premature discharges and engage in evidence-based discussions with COVID-19 patients
regarding discharge plans. It may also reduce strain on hospital capacity by identifying patients safe for
discharge and reserving resources for higher-risk COVID-19 patients.
The project will be executed by a multidisciplinary team with expertise in emergency care, quality outcomes
research, care transitions, and applying data science to improve clinical care, including ML and NLP methods.
It will also use innovative methods, including a mixed methods approach to iteratively develop the concept map
that will inform the predictive model. Moreover, the proposed project is designed to optimize the generalizability
of CERST, by using a large, diverse study population, including data from a second health system with a
different EHR using Fast Health Interoperability Resources (FHIR) specifications to assist with model
interoperability. This will help optimize model performance for differing patient populations, health systems, and
EHR platforms. Since the primary data source for this study is readily accessible to the study team, who
possesses prior experience working with the data sources and performing the analytic procedures outlined in
the proposal, the team is well-positioned to execute this study with timely dissemination of project findings.
